OpenAI brings full-duplex GPT-Live-1 voice model to its API

OpenAI has launched GPT-Live-1 in the API, bringing the company’s simultaneous listening-and-speaking voice technology to developers. In its OpenAI News announcement dated September 10, 2026, the company says the model is designed for more natural voice applications and business workflows.

Unlike a conventional speech-to-text, language-model, and text-to-speech pipeline, GPT-Live-1 processes incoming and outgoing audio together. That allows it to respond more smoothly when users interrupt, pause, change direction, or speak over an answer. OpenAI says early testing by language-learning company Speak found nearly 80% fewer interruptions than with an earlier turn-based system.

More control for developers

Builders can steer an agent’s tone, pace, and style through the system prompt, while delegating complex reasoning and tool calls to a backend text model or third-party model. The API also includes transcripts, response text, keyword biasing, turn detection, improved handling of silence and background noise, and better context retention for longer sessions.

The release is aimed at use cases ranging from reservations and customer support to telephony. OpenAI says one partner reduced its voice-agent codebase by 80%, removing 23,000 lines of code, though that result comes from the partner’s own comparison.

OpenAI reports a 30-point improvement over GPT-Realtime-2.1 on Full Duplex Bench performance. Its reported turn-taking latency was 0.798 seconds, compared with 1.41 seconds for GPT-Realtime-2.1 and 1.63 seconds for GPT-Realtime-2. It also scored 86.2% on Tau3 Voice when paired with GPT-6 Astra at medium reasoning effort, versus 45.7% and 42.4% for the two comparison models.

For AI developers, the significance is architectural: voice agents can remain responsive in real time while heavier reasoning and external actions happen behind the scenes.

Source: OpenAI News